2026-04-08
By Vadym · Generated with AI, curated by me
Security went from afterthought to proving ground overnight. Anthropic unveiled a model that finds zero-days at superhuman scale — and kept it in a locked vault. A 26-person startup answered by shipping one of the largest open-weight reasoning models ever released. The frontier is dividing along a single axis: controlled access or radical openness. Both sides moved forward today.
Anthropic announced Project Glasswing alongside a preview of Claude Mythos, a frontier model capable of discovering software vulnerabilities at a level surpassing all but the most skilled human security researchers. Mythos has already identified thousands of high-severity zero-days across major operating systems and browsers. Rather than releasing it publicly, Anthropic is granting controlled access to 12 founding partners — including AWS, Apple, Cisco, CrowdStrike, JPMorganChase, Linux Foundation, Microsoft, and NVIDIA — backed by $100M in usage credits and $4M in open-source security donations.
Arcee, a 26-person U.S. startup, released Trinity Large Thinking: a 399-billion parameter open-source reasoning model under Apache 2.0. It uses a sparse Mixture-of-Experts architecture (~13B active parameters per token) trained over 33 days on 2,048 NVIDIA B300 Blackwell GPUs at a cost of $20M — nearly half the company’s total funding. CEO Mark McQuade claims it is the most capable open-weight reasoning model released by a non-Chinese company, explicitly framing it as a domestically-controlled alternative to Chinese-built LLMs for Western enterprises.
GitHub expanded Copilot’s cloud agent into a full planning and research loop. New capabilities: branch-based work without requiring a pull request, implementation plan review before any code is written, and deep codebase research sessions that build context before execution. The same week, GitHub released the Copilot SDK in public preview (April 2), letting developers embed Copilot’s agentic capabilities in external applications, and added bring-your-own-key support for local model integration.
Meta is developing two new proprietary frontier models: “Avocado,” a large language model, and “Mango,” a multimedia file generator, both expected to launch in 2026. Meta plans to release open-source derivatives of each after the proprietary versions ship — but with certain capabilities removed, particularly in cybersecurity-adjacent code generation. Meta acknowledges the models may not outperform competitors across all benchmarks but sees advantages in specific domains.
OpenAI published a policy document outlining an economic framework for the “intelligence age.” Three core proposals: sovereign public wealth funds to distribute AI-driven prosperity broadly; a robot and automation tax to cushion workforce displacement; and a four-day workweek as labor adapts to AI-augmented productivity. The document acknowledges AI as a displacement force while arguing the gains can be shared rather than concentrated — a notable rhetorical shift from pure capability-first messaging.
Intel has signed on to Terafab, a U.S. semiconductor manufacturing project in Texas anchored by SpaceX and Tesla. The specific scope of Intel’s contributions has not been publicly disclosed. The project is one of several initiatives aimed at expanding domestic chip production capacity, combining Intel’s foundry expertise with SpaceX’s manufacturing scale and Tesla’s supply chain experience.
Today’s stories run a single thread: what gets released versus what gets locked. Anthropic shipped a model that outperforms most security researchers at finding vulnerabilities — and gave it only to 12 partners. Meta pre-announced open-source derivatives of its next two models and immediately specified what gets removed. GitHub’s cloud agent now researches your codebase before writing a line. OpenAI published an economic manifesto that acknowledges displacement while targeting a trillion-dollar valuation. And Intel quietly joined a Texas chip project that nobody’s explaining in detail. The debate about who gets access to the most capable AI is no longer theoretical. It’s a design decision being made right now, model by model, partner list by partner list.
— Boba
Curated by Vadym